The Potential Performance of Artificial Neural Networks in QSTRs for Predicting Ecotoxicity of Environmental Pollutants

This review surveys the applications of neural network methodologies to the field of Quantitative Structure-Toxicity Relationships (QSTRs) in environment, and more specifically ecotoxicity. QSTR is one of the methods for predicting hazards of various chemicals and utilizes a computer-based technology such as artificial neural network to predict the toxicity of a chemical solely from its molecular attributes. Many artificial neural network methodologies have been applied to ecotoxicological data for fish, bacteria, protozoa and so on. The results demonstrate the ability of the artificial neural network methodologies to apply nonlinear structure-toxicity relationships for the prediction of the corresponding toxicity values for chemicals, which are not part of the training sets. In order to employ an artificial neural network for QSTR, although users must pay attention to over-parameterization, data distribution, the structure and training cycle of neural network, and chance correlation, fine tuned neural network has high performance to predict ecotoxicity of chemicals. In the most of the QSTR studies, the results by artificial neural network modeling gave clearly better prediction of toxicity values compared to the results by multiple linear regression analysis or other commercial QSTR programs.